Simulator
The Simulator (simulator) is a built-in, in-process fake agent. It needs no installation and no API key, and it’s always detected as available (version built-in). The Demo: Issue to PR (simulated) pipeline created on first launch uses it for every agent step.
Use it to:
- learn the app and watch a full run (parallel steps, a review loop, an approval gate) without spending anything;
- dry-run a new pipeline’s structure, dependencies, loops,
runIfconditions and approvals before switching steps to real harnesses; - try Generate with AI in the step editor without a real agent.
The Simulator doesn’t touch files. It only writes log lines and returns canned output. It has no Configure button and can’t be reset or deleted, but it can be disabled.
Models
Section titled “Models”The model controls its behaviour:
| Model | Behaviour |
|---|---|
fast |
Roughly a quarter of the delays. |
realistic (default) |
Multi-second run with simulated tool calls. Reviewers request changes on the first pass. |
flaky |
Fails on the first attempt (Simulated transient failure), succeeds on retry. Use it to test Retries on failure. |
What it outputs
Section titled “What it outputs”The Simulator looks at the rendered prompt to decide what kind of step it’s playing:
| If the prompt… | It behaves like | Output |
|---|---|---|
contains VERDICT |
a reviewer | With realistic on loop iteration 0: two review comments and VERDICT: CHANGES_REQUESTED. Otherwise: VERDICT: APPROVE. |
starts with a line matching planning agent or plan |
a planner | A short Markdown plan. |
starts with a line containing classify or triage |
a triager | Type: feature · Size: S · Risk: low · Area: api |
| anything else | an implementer | Implemented the change… plus bullets. On loop iterations it adds (revision N, addressed review feedback). |
So a review step with Pass if output matches (regex) VERDICT:\s*APPROVE and On failure, loop back to set will loop exactly once with the realistic model, which is how the demo shows the feedback loop.
It also logs fake tool calls (Glob, Read, Edit, Bash: npm test, …), a thinking line, and a result line. It reports made-up token counts and a cost computed from them at $3 / $15 per million input / output tokens, so the Dashboard’s spend and token figures have something to show.
Limits
Section titled “Limits”- The Simulator can’t design pipelines. Build with AI in the New pipeline dialog rejects it with The simulator cannot design pipelines. Choose a real harness.
- For Generate with AI on a step it returns canned text: a generic prompt template, a generic system prompt, or
npm ci && npm testfor a shell command. - Shell steps and approval steps in a Simulator pipeline are real. The demo’s shell steps only
echoandsleep.